Online Anomaly Detection and Adaptive Method for Multivariate Profile Data in Wafer Manufacturing

To address the dual requirements of high real-time performance and robust reliability in quality monitoring during the semiconductor wafer manufacturing process, this work proposes an online anomaly detection and adaptive method for multivariate profile data. First, an anomaly detection framework based on principal component analysis (PCA) and Hotelling T² statistic is constructed, realizing early anomaly identification through feature dimensionality reduction and multivariate statistical monitoring. Second, sliding window and parameter adaptive update mechanism are introduced to dynamically compensate for long-term data drift induced by equipment aging,environmental fluctuations and other factors, thereby enhancing the stability and adaptability of the model in continuous production. Experiments on a wafer manufacturing dataset containing 52 samples, including 6 anomalous samples, show that the proposed method achieves a detection accuracy of 96.15%, an F₁-score of 85.71%, with a detection delay of only 1 sample and an average processing time per sample of less than 0.01 second. Compared with traditional control charts, it significantly reduces the false alarm rate while maintaining high detection accuracy. In addition, the scalability and robustness of the proposed framework are further verified by constructing a tensor regression comparison model. This study offers a highly sensitive and adaptive solution for real-time quality monitoring in semiconductor manufacturing, possessing great prospects for practical engineering application.